TL;DR
This paper develops methods for AI agents to strategically seek information and make decisions, inspired by human cognition, improving their performance in tasks like Battleship and Guess Who? with novel inference strategies.
Contribution
The paper introduces a decision-oriented dialogue task and Monte Carlo inference strategies inspired by Bayesian Experimental Design to enhance AI agents' information-seeking abilities.
Findings
LM agents struggle with asking informative questions and accurate answers
Monte Carlo inference boosts Spotter accuracy by up to 14.7%
Methods enable weaker LMs to outperform humans and frontier models in accuracy and win rates
Abstract
Many emerging applications of AI--from scientific discovery to medical diagnosis--require agents to seek information strategically: forming hypotheses, asking targeted questions, and making decisions under uncertainty. In high-stakes settings with limited resources, do language models (LMs) behave like rational agents? Drawing on insights from human cognition, we develop methods to evaluate and enhance agentic information-seeking. First, we introduce a decision-oriented dialogue task called Collaborative Battleship, in which a Captain must balance exploration (asking questions) and action (taking shots), while a Spotter must supply accurate, contextually-grounded answers. Compared to human players (N=42), we find that many LM agents struggle to ask informative questions, produce accurate answers, and identify high-utility actions. To address these gaps, we develop novel Monte Carlo…
Peer Reviews
Decision·ICLR 2026 Oral
There are numerous strengths with this paper. To start, the collaborative battleship game sets up an interesting scenario for the captain agent, where they not only have to balance information seeking vs. reward seeking behaviour, but also maintain uncertainty about the correctness of its collaborator (noisy spotter agent). Information seeking in the space of natural language is an interesting setting to study. The human dataset provides valuable information to ground “human level” performance,
A number of simplifying assumptions were made in the “bayes rational” modelling choices. For instance, modelling with fixed $\epsilon$ (as the authors already point out) and $\gamma$ (not sure how this is set), and only modelling single-step look-ahead are simplifying assumptions. I do not think this detracts from the main point of the paper, as to my understanding this paper is about improving empirical performance of weak LMs using some cognitively inspired “bayes rational" strategies. Neverth
**Novel Task and Dataset:** Collaborative Battleship provides a clean, interpretable environment for studying the explore/exploit dilemma and grounded communication. The BATTLESHIPQA dataset, collected from human interactions, is a valuable contribution for benchmarking and analysis. **Principled Bayesian Framework:** The application of BED principles (EIG maximization, belief updating via SMC, MAP action selection) provides a strong theoretical grounding for the proposed inference-time strateg
**Limited Scope:** While the paper introduces an interesting method, its evaluation is confined to a specific domain defined by the authors, with generalization demonstrated only on one additional ad-hoc task. This limited scope makes it difficult to assess the method's potential for broader applicability and overall impact. **"Surpass Humans" Claim Qualification:** The claim that augmented LMs surpass human performance needs stronger qualification. Details on human participants' prior experien
I enjoyed this paper a lot. It has all the ingredients for a great paper: a new task, human evaluation, a decent set of different models that are evaluated, a novel technique for improving models, and demonstration that the findings generalize to another domain. The paper was exceptionally well written and easy to follow. The method is clean and simple, yet effective. CaptainQA is an interesting agentic test bed for LLMs and the whole methodology fits thematically well into ICLR.
I found the usage of the term Bayes-rational weird. Best to my knowledge, this is not an accepted term in the literature. It implies that the strategies developed by the authors are Bayes-optimal, which is not the case (as also noted by the authors). To avoid this confusion, I would suggest using a different term instead. There is not so much negative to say about the paper. Perhaps the only downside is that, while the results and methods are interesting, they are not groundbreaking. For me, th
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
